gzivdo/fcpe-onnx

FCPE ONNX — unofficial export

Pre-converted ONNX export of FCPE (Fast Context-based Pitch Estimation, CN_ChiTu, arXiv 2509.15140).

This is an unofficial community export of the bundled torchfcpe checkpoint, intended for use without the PyTorch dependency. The weights and architecture are unchanged — only the runtime is swapped from torch to ONNX Runtime.

Provenance

• Upstream code & weights: https://github.com/CNChTu/FCPE (MIT — see LICENSE)
• Upstream paper: Tu, "FCPE: A Fast Context-based Pitch Estimation Model", arXiv 2509.15140, 2025
• Bundled checkpoint version: torchfcpe == 0.0.4 (PyPI)
• Export script (this conversion): pitch-core/tools/fcpe_export.py (MIT OR Apache-2.0, copyright 2026 gzivdo)
• Reproduction: python tools/fcpe_export.py --out fcpe.onnx (requires pip install torch torchfcpe)

This export is not endorsed by, affiliated with, or sponsored by the FCPE authors. It is provided as a convenience for the open-source community.

I/O contract

input:  audio   float32  [1, n_samples, 1]   raw mono audio @ 16 kHz
output: f0_hz   float32  [1, n_frames, 1]    f0 in Hz (0 = unvoiced)

• Sample rate: 16 000 Hz (resample your input before feeding)
• Hop: 160 samples = 10 ms
• Output frames: n_samples // 160 + 1
• Voicing gate: model applies internal threshold=0.006 on confidence; frames with confidence below it are returned as f0=0. Some quiet frames may also return NaN (internal log(0)) — treat as unvoiced.

Usage (Python)

import numpy as np
import onnxruntime as ort
import librosa

audio, _ = librosa.load("vocal.wav", sr=16_000, mono=True)
sess = ort.InferenceSession("fcpe.onnx", providers=["CPUExecutionProvider"])
f0 = sess.run(["f0_hz"], {"audio": audio.astype(np.float32)[None, :, None]})[0]
f0 = f0[0, :, 0]
voiced = np.isfinite(f0) & (f0 > 0)
print(f"voiced: {voiced.sum()}/{len(f0)} frames")

Usage (Rust via pitch-core-onnx)

use pitch_core::PitchTracker;
use pitch_core_onnx::FcpeEstimator;

let est = FcpeEstimator::new("fcpe.onnx")?;
let mut tracker = PitchTracker::new(est, 48_000, 1024)?;
for frame in tracker.process(&audio_chunk)? { /* ... */ }

See https://crates.io/crates/pitch-core-onnx for the full crate.

Citation

If you use this model in academic work, cite the upstream paper, not this export:

@article{tu2025fcpe,
  title  = {FCPE: A Fast Context-based Pitch Estimation Model},
  author = {CN\_ChiTu},
  journal = {arXiv preprint arXiv:2509.15140},
  year   = {2025},
  url    = {https://arxiv.org/abs/2509.15140}
}

License

This ONNX file inherits the MIT license from the FCPE upstream:

MIT License

Copyright (c) 2023 CN_ChiTu

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction […]

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY […]

Full text: https://github.com/CNChTu/FCPE/blob/main/LICENSE

Disclaimer

The export script tools/fcpe_export.py applies a small monkey-patch to torch.stft so the legacy ONNX tracer can handle the complex-typed output from torchfcpe's mel extractor. The patch wraps the real-tensor output in a _FakeComplex shim that exposes .real / .imag as indexed views — semantically equivalent to the original. Numerical output should match the upstream torchfcpe model bit-for-bit modulo floating-point rounding in the ORT runtime.

This file is provided "AS IS", per the MIT license above. The maintainer makes no claims about its accuracy on data outside the ranges tested by upstream and provides no warranty of fitness for any particular purpose.

If the upstream FCPE project releases an official ONNX export, prefer that. If you find a discrepancy between this export and upstream torchfcpe inference, please open an issue at https://github.com/gzivdo/pitch-core/issues.